Closed thekingofkings closed 7 years ago
The leaveOneOut
evaluation in KDD submission has minor issues. It normalizes generates all features first, and then leave the testing sample out. In this case, the geo
and taxi
features for training samples actually contain information about the crime in testing sample.
normalization by destination
Known bug in #2 #8.
The transpose
is missing. We need transpose, because we are multiplying flow matrix
with the crime vector
.
There is a table "comparing different feature settings for overall crime" in KDD paper. Now the experiment setting to get this table is not clear.
Download the
release v2.0
to figure this out.